Abstract
As ubiquitous wearable sensors capture a wealth of behavioral and physiological information, transforming these signals into personalized health insights remains challenging. Specifically, converting low-level sensor data into representations that can characterize higher-level states is difficult due to high phenotypic diversity and variations in individual baseline health, physiology, and lifestyle factors.
Moreover, collecting wearable data paired with health outcome annotations is labor-intensive and costly, while retrospective annotation is practically unfeasible, contributing to a scarcity of high-quality labeled data. To address these limitations, we propose a foundation model for wearable health pretrained on over one trillion minutes of unlabeled sensor signals from a large cohort of five million participants.
We demonstrate that the joint scaling of model capacity and pretraining data volume leads to systematic performance improvements across a diverse set of 35 health prediction tasks, including cardiovascular, metabolic, sleep, and mental health, as well as lifestyle choices and demographic factors. We find that this population-scale representation enables label-efficient few-shot learning and generative capabilities for robust daily metric estimation.
Furthermore, we deploy a classroom of LLM agents to autonomously explore the space of downstream predictive heads built on the model embeddings, showing broad performance enhancements that increase with LLM model capacity. Finally, we demonstrate how integrating these downstream predictors into a Personal Health Agent can support model responses that are more relevant, contextually aware, and safe, validated through 1,860 ratings from a cohort of clinicians.
Blogger's Review: This paper showcases the immense potential of foundation models in the wearable health data domain, particularly through innovative methods to address data scarcity via large-scale pretraining. The integration of LLM with health predictions offers a fresh perspective and practical direction for future personalized health management.